If this is a bulk import report, send the spreadsheet to services@wildme.org with the email subject line matching your bug report
In which Wildbook did the issue occur? Internet of Turtles
What operating system were you using? (eg. MacOS 10.15.3)
What web browser were you using? (eg. Chrome 79)
What is your role on the site? (admin, researcher, etc) researcher
What happened?
Used bulk import to upload ~15 photos, worked great. The import was sent to detection and when complete the summary (under the table) showed
Summary
Total marked individuals: 0
Total encounters: 13
Total media assets: 15
- Number with acmIDs: 15
- Number valid for image analysis: 15
- Number that have completed detection: 15
Total annotations: 11 (11 matchable)
- {turtle_green=right}: 7
- {turtle_green=left}: 4
Images sent to IA: 15 (100%)
mage formats generated? *yes
What did you expect to happen?
Prior imports, once through detection, show the results in the import table above the Summary.
What are some steps we could take to reproduce the issue? https://iot.wildbook.org/import.jsp?taskId=38c18ae5-3a4e-4d69-8974-c888ddf4f9eb
Hi @GoBeach
The table status looks correct. It shows all of the Match Results by Class as completed. However, it does look like the annotations themselves are incorrect as they’re not selecting the full head. I’ll need more time to look into the inaccurate detections.
A better table would be this one: https://iot.wildbook.org/import.jsp?taskId=a828cb74-f899-4c17-ba04-0b3617cae0e5 right after I created the post I saw the column populate.
1 Like
Great; thanks! I hope to have an update soon.
I have another big upload ready to go - should I wait?
I just met with the devs and learned that the reason for both the missing annotations and the incorrect ones have the same root cause: the algorithm was never trained on partially submerged green turtles. The fact that it’s detecting any without being trained on it is a good sign.
It does mean that you’ll need to manually annotate any that didn’t get detected or delete and re-draw annotations that aren’t giving you appropriate match candidates. Once you’ve uploaded and identified enough partially submerged green turtles, we can revisit re-training the algorithm with these images.
You can still upload new bulk imports. Because these are new perspectives for the match algorithms, it’ll need a little more hands-on intervention from you/your team with the guidance above to prepare the data for retraining efforts later.
I’m not sure I agree with that 100%. The smaller uploads I’ve done have had some annotations, but the large upload didn’t get many. Maybe I was looking at it wrong - I’ll submit this one I have ready and we’ll see what happens.
To clarify, all it means that the lack of training data with this perspective means that you’ll see a higher number of overall missed/partial annotations compared to images of turtles that are completely underwater or completely on land.
I uploaded a large set of data over the weekend, the summary says there were 26 annotations, but I don’t see any of them noted on the table: https://iot.wildbook.org/import.jsp?taskId=5ca7c4ce-05c6-4b50-9be1-ab99bbe88fac
I’m waiting to hear back from the devs on this one. I clicked on one of the encounters in the table and did find an annotation for it despite it not reflecting in the table: Internet of Turtles
Thanks for hanging in there. I double-checked with the devs and this behavior is normal. The IA classes won’t appear in the column until you start ID. As soon as I clicked the Send to identification button, the IA classes for the annotated encounters displayed right away.